@archsight/aios 1.3.2 → 1.5.0

package/docs/v1.5.0-knowledge-pack-runtime.md

@@ -0,0 +1,100 @@
+# v1.5.0 Knowledge Pack 与本地 Reference Runtime
+
+v1.5.0 的主线是把 AIOS 从“能写工程文档”推进到“能治理工程知识”。核心交付不是新增一批松散 Prompt，而是引入 AIOS 原生知识资产：`Knowledge Pack`。
+
+## 版本定位
+
+Knowledge Pack 是可编译、可查询、可评估的工程知识资产。它把规范摘录、企业标准、项目资料、历史审查口径和人工复核记录治理为一个统一对象，再交给 Skill、RAG / GraphRAG 流程和 Capability Runtime 使用。
+
+本版本同时提供本地 Reference Runtime：`knowledge.norm_lookup` 可以查询编译后的 Knowledge Pack，返回 `status`、`citations`、`applicability`、`sourceVersion` 和 `notes`，并按 Capability 仲裁规则决定是否 `proceed`、`hold` 或升级人工复核。
+
+## Knowledge Pack 文件链
+
+```text
+knowledge-pack.source.json
+source-register.md
+standard-register.md
+clause-map.md
+entity-relation-map.md
+eval-questions.md
+review-notes.md
+        |
+        v
+compiled/knowledge-pack.json
+        |
+        v
+knowledge:lookup / knowledge:eval / knowledge.norm_lookup
+```
+
+`knowledge-pack.source.json` 是结构化源，Markdown 文件用于业务人员阅读、复核和交接。编译产物 `compiled/knowledge-pack.json` 是 Runtime 消费对象。
+
+## CLI
+
+```bash
+archsight-aios knowledge:init --sample --name scheme-review
+archsight-aios knowledge:validate --name scheme-review
+archsight-aios knowledge:compile --name scheme-review
+archsight-aios knowledge:inspect --name scheme-review
+archsight-aios knowledge:lookup --pack scheme-review/compiled/knowledge-pack.json --query "高支模方案是否应检查计算书"
+archsight-aios knowledge:eval --name scheme-review
+```
+
+发布验证入口：
+
+```bash
+npm run validate:knowledge-pack
+```
+
+该命令会验证样板工作台、编译 Knowledge Pack、检查 inspect 摘要、执行 lookup、跑 eval gate，并通过 `capability:call` 验证本地 `knowledge.norm_lookup` MCP Reference Runtime。
+
+## Reference Runtime 结果契约
+
+```json
+{
+  "status": "found",
+  "citations": [
+    {
+      "sourceId": "SRC-SYN-SCHEME",
+      "standardId": "STD-SYN-SCHEME",
+      "sourceVersion": "SYN-2026.01",
+      "clauseId": "CLAUSE-HIGH-FORMWORK",
+      "clauseNo": "S-2.1",
+      "title": "高支模方案计算书要求",
+      "pageRange": "p.7",
+      "reviewStatus": "reviewed"
+    }
+  ],
+  "applicability": "applicable",
+  "sourceVersion": "SYN-2026.01",
+  "notes": "命中高支模计算书口径；缺计算书时不得判断方案通过。"
+}
+```
+
+`status` 支持：
+
+- `found`：命中 Knowledge Pack。
+- `not_found`：未命中，不得用模型常识补成规范结论。
+- `conflict`：命中冲突口径，应阻断自动结论。
+- `inapplicable`：版本、地区、专业或条件不适用。
+- `error`：Runtime 或数据错误。
+
+`applicability` 支持：
+
+- `applicable`
+- `not_applicable`
+- `need_context`
+
+## 与 Skill 的关系
+
+- `aios-knowledge`：生产和维护 Knowledge Pack。
+- `aios-runtime`：设计 Knowledge Pack 进入 RAG / GraphRAG / Tool Calling 的方式。
+- `aios-review`：审查引用、版本、越权结论和 eval gate。
+- `aios-scheme-audit`：在专项施工方案复核中优先使用 Knowledge Pack 和 `knowledge.norm_lookup` 证据；缺工具证据时标注 `需核验`。已使用 `aios-construction-scheme` 的团队可继续使用原领域型入口。
+
+## 边界
+
+- v1.5.0 不提供生产级法规 / 规范数据库。
+- v1.5.0 不承诺自动合规判定。
+- 未授权规范全文不得进入仓库。
+- 合成样板只验证知识治理结构和 Runtime 契约，不代表真实规范条文。
+- `published` 只表示包内资料通过本项目复核，不替代法务、总工、专家、审图或注册人员签审。

package/gemini-extension.json

@@ -1,6 +1,6 @@
 {
   "name": "archsight-aios",
-  "version": "1.3.2",
+  "version": "1.5.0",
   "description": "面向建筑行业知识工作从业者与 AI 研发团队的 Skills、Workflow 与多 Agent 工具包 / Building-industry AI agent skills for BIM, IFC, RAG, GraphRAG, project evidence work, code review, and runtime governance.",
   "contextFileName": "GEMINI.md"
 }

package/knowledge/README.md

@@ -1,6 +1,6 @@
 # Knowledge
 
-`knowledge/` 保存建筑行业知识体系。
+`knowledge/` 保存建筑行业知识体系和治理政策。v1.5.0 后，真正可运行的知识资产以 `Knowledge Pack` 为核心对象，模板和样板位于 `templates/knowledge-pack/` 与 `templates/knowledge-pack-samples/`。
 
 推荐方向：
 
@@ -12,9 +12,12 @@
 - 结构知识
 - 审查规则
 
-该目录未来应支撑 GraphRAG、知识图谱、条文结构化和智能审查。
-
 当前入口：
 
 - [领域分类](domain-taxonomy.md)
 - [知识来源登记](source-register.md)
+- [v1.5.0 Knowledge Pack 与本地 Reference Runtime](../docs/v1.5.0-knowledge-pack-runtime.md)
+
+## Knowledge Pack 关系
+
+Knowledge Pack 把来源、版本、条文、实体、关系、查询规则、评估问题和人工复核状态编译成 `archsight-aios.knowledge-pack`，再交给 `knowledge.norm_lookup`、RAG / GraphRAG 管线和专项 Skill 使用。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@archsight/aios",
-  "version": "1.3.2",
+  "version": "1.5.0",
   "description": "面向建筑行业知识工作从业者与 AI 研发团队的 Skills、Workflow 与多 Agent 工具包 / Building-industry AI agent skills for BIM, IFC, RAG, GraphRAG, project evidence work, code review, and runtime governance.",
   "type": "module",
   "homepage": "https://github.com/ArchSightLabs/archsight-aios#readme",
@@ -23,7 +23,11 @@
     "construction-ai",
     "building-ai",
     "aec",
+    "tender",
+    "technical-bid",
+    "scheme-audit",
     "building-code",
+    "knowledge-pack",
     "project-evidence-work",
     "agent-skills",
     "ai-agent",
@@ -51,11 +55,16 @@
     "validate:prompts": "node ./scripts/validate-prompt-fixtures.mjs",
     "validate:prompt-run-pack": "node ./scripts/build-prompt-run-pack.mjs --check",
     "validate:public-advisory-run-pack": "node ./scripts/build-prompt-run-pack.mjs --fixture prompts/evaluations/engineering-business-public-advisory-fixtures.json --check",
+    "validate:document-writing-run-pack": "node ./scripts/build-prompt-run-pack.mjs --fixture prompts/evaluations/engineering-document-writing-fixtures.json --check",
     "validate:prompt-run-results": "node ./scripts/validate-prompt-run-results.mjs --check-template",
     "validate:prompt-outputs": "node ./scripts/validate-prompt-model-outputs.mjs",
     "validate:prompt-scorecard": "node ./scripts/validate-prompt-scorecard.mjs",
+    "validate:document-writing-scorecard": "node ./scripts/validate-prompt-scorecard.mjs",
+    "validate:skill-runtime-evidence": "node ./scripts/validate-skill-runtime-evidence.mjs",
+    "validate:knowledge-pack": "node ./scripts/validate-knowledge-pack.mjs",
     "build:prompt-run-pack": "node ./scripts/build-prompt-run-pack.mjs --out prompts/evaluations/engineering-business-basic-run-pack.generated.json",
     "build:public-advisory-run-pack": "node ./scripts/build-prompt-run-pack.mjs --fixture prompts/evaluations/engineering-business-public-advisory-fixtures.json --out prompts/evaluations/engineering-business-public-advisory-run-pack.generated.json",
+    "build:document-writing-run-pack": "node ./scripts/build-prompt-run-pack.mjs --fixture prompts/evaluations/engineering-document-writing-fixtures.json --out prompts/evaluations/engineering-document-writing-run-pack.generated.json",
     "analyze:prompt-run-results": "node ./scripts/analyze-prompt-run-results.mjs",
     "test": "node ./tests/cli.test.mjs"
   },

package/prompts/README.md

@@ -16,7 +16,11 @@ Prompt 会腐化，因此不能只保存文本本身，必须保存评估和维
 对比验证记录见 [工程业务管理基础提示词对比验证](evaluations/engineering-business-basic-prompts-2026-06-16.md)。
 advisory 来源信号复核见 [工程业务基础提示词 advisory 复核说明](evaluations/engineering-business-basic-advisory-validation-2026-06-16.md)。
 结构化评分卡见 [工程业务管理基础提示词评分卡](evaluations/engineering-business-basic-scorecard.json)，可用 `npm run validate:prompt-scorecard` 校验。
+写作型评分卡见 [工程文档写作评分卡](evaluations/engineering-document-writing-scorecard.json)，同样由 `npm run validate:prompt-scorecard` 或 `npm run validate:document-writing-scorecard` 校验。
+Knowledge Pack 评分卡见 [工程知识包评分卡](evaluations/engineering-knowledge-pack-scorecard.json)，由 `npm run validate:knowledge-pack` 校验，并覆盖来源治理、版本适用性、Runtime 查询和 eval gate。
 weak/basic 运行包可用 `npm run build:prompt-run-pack` 生成，生成前可用 `npm run validate:prompt-run-pack` 校验。
+写作型 Skill 运行包使用 `npm run build:document-writing-run-pack` 生成，生成前可用 `npm run validate:document-writing-run-pack` 校验。
+真实宿主触发证据见 [Skill Runtime Evidence](evaluations/skill-runtime/)，可用 `npm run validate:skill-runtime-evidence` 校验归档结构。
 weak/basic 运行结果模板可用 `node ./scripts/validate-prompt-run-results.mjs --init <file>` 生成，模板和真实结果可用 `npm run validate:prompt-run-results` 或 `--file` 校验。
 weak/basic 运行结果报告可用 `npm run analyze:prompt-run-results -- --file <results> --out <report>` 生成。
 输出结构样例见 [工程业务管理基础模型输出样例](evaluations/engineering-business-basic-model-output.example.json)，可用 `npm run validate:prompt-outputs` 校验。

package/prompts/evaluation-policy.md

@@ -30,6 +30,39 @@ npm run validate:prompts
 
 该检查不替代真实模型输出评估，但能保证 6 类基础场景、抽象来源信号、必备输出结构、禁止结论和敏感信息边界没有被破坏。
 
+## 工程文档写作提示词回归
+
+工程文档写作型提示词使用 `prompts/evaluations/engineering-document-writing-fixtures.json` 作为脱敏回归基线，当前覆盖 `aios-tender-write` 和 `aios-scheme-write` 两类生成 / 改写任务；`aios-contract-draft`、`aios-daily-write` 和 `aios-meeting-write` 先按 Skill 结构测试纳入边界，后续可补充独立 fixture。
+
+该 fixture 不验证模型真实文采，也不证明可以直接交付客户；它只检查写作型 Skill 是否稳定保留：
+
+- Markdown 工作母版。
+- 资料来源、写作 brief 和历史素材复用判断。
+- 章节初稿 / 改写稿和待补占位。
+- 禁止编造、禁止越权结论和人工复核边界。
+- 生成后交回对应审核门禁：`aios-tender-audit`、`aios-contract-audit`、`aios-construction-daily`、`aios-construction-meeting` 或 `aios-scheme-audit`；历史流程中的 `aios-commercial-tender`、`aios-commercial-contract`、`aios-construction-daily`、`aios-construction-meeting` 和 `aios-construction-scheme` 继续保留。
+
+若需要批量运行写作 weak/basic 对照输入，使用：
+
+```bash
+npm run validate:document-writing-run-pack
+npm run build:document-writing-run-pack
+```
+
+写作运行包包含 2 个 case 的普通提示词和基础提示词两组输入，共 4 条 run item。该步骤只组织脱敏 / 虚构输入和 prompt 文本，不调用模型。
+
+写作型评分卡保存在 `prompts/evaluations/engineering-document-writing-scorecard.json`，用于固定来源链、历史素材复用、初稿可操作性、边界安全、审核门禁和交接可读性 6 个维度。修改写作 fixture、写作 prompt 或评分维度后，运行：
+
+```bash
+npm run validate:document-writing-scorecard
+```
+
+真实宿主触发证据保存在 `prompts/evaluations/skill-runtime/`。该目录允许记录 blocked / uncertain 状态，但不能把 blocked case 写成通过；只有原始输出已归档且能确认真实触发对应 Skill 时，才允许标记 `skillRuntimeConfirmed=yes`。
+
+```bash
+npm run validate:skill-runtime-evidence
+```
+
 普通提示词与基础提示词的结构化比较保存在 `prompts/evaluations/engineering-business-basic-scorecard.json`。修改 fixture、基础提示词或评分维度后，运行：
 
 ```bash

package/prompts/evaluations/engineering-document-writing-fixtures.json

@@ -0,0 +1,129 @@
+{
+  "schema": 1,
+  "name": "engineering-document-writing-fixtures",
+  "version": "0.1",
+  "dataBoundary": "De-identified engineering document writing fixtures. Customer names, people, locations, dates, amounts, project numbers, and raw source documents are synthetic or abstracted.",
+  "sourceBoundary": "Source signals preserve writing-task shape only: source normalization, historical material reuse, draft generation, and review-gate handoff.",
+  "cases": [
+    {
+      "id": "tender-writing-historical-material-reuse",
+      "skillId": "aios-tender-write",
+      "promptPath": "skills/aios-tender-write/prompts/basic-prompt.md",
+      "scenario": "基于招标要求、评分点、用户初稿和历史标书素材生成技术标章节工作母版",
+      "sourceSignals": [
+        "source-shape:technical bid writing task with scoring points and historical material snippets",
+        "source-shape:user draft needs rewriting into a markdown working draft",
+        "boundary-shape:historical project facts cannot be copied into the new project without evidence"
+      ],
+      "advisoryComparison": [
+        "普通提示词容易直接扩写成顺滑正文，忽略来源链和本项目事实缺口。",
+        "写作型基础提示词要求先建立资料来源、写作 brief、素材复用判断和审核门禁。"
+      ],
+      "weakPrompt": "请根据这些招标要求和历史标书，帮我写一版技术标章节。",
+      "inputSummary": [
+        "输入包含技术标要求、评分点摘要、用户初稿和历史标书片段。",
+        "历史标书含有旧项目名称、工期、人员和设备表述，需要先判断是否可复用。",
+        "缺少本项目图纸、组织机构、资源计划和证明材料时，只能输出待补占位。"
+      ],
+      "sampleInput": [
+        "资料类型：技术标章节生成任务。",
+        "招标要求摘要：需说明施工总体部署、进度计划、质量安全文明施工和资源配置；评分点要求章节与措施一一对应。",
+        "用户初稿：希望复用历史项目中的总体部署和安全文明施工段落，但本项目施工组织、设备、人员和工期尚未确认。",
+        "历史素材片段：旧项目写有项目名称、管理人员、设备型号、工期安排和现场布置描述。",
+        "资料缺口：未提供招标原文页码、图纸、项目组织架构、资源计划、证明材料和最终投标格式要求。"
+      ],
+      "requiredPromptTerms": [
+        "Markdown",
+        "复用级别",
+        "评分点到章节映射",
+        "审核门禁交接",
+        "aios-commercial-tender"
+      ],
+      "expectedOutputShape": [
+        "先列资料来源和写作 brief。",
+        "把历史素材按可复用改写、仅参考、不可套用、需人工确认分级。",
+        "输出评分点到章节映射和 Markdown 章节初稿。",
+        "缺少依据时写待补占位，并交给审核 Skill 复核。"
+      ],
+      "expectedStrongSections": [
+        "资料来源清单",
+        "写作 brief",
+        "历史素材匹配表",
+        "评分点到章节映射",
+        "章节初稿 / 改写稿",
+        "审核门禁交接"
+      ],
+      "weakFailureModes": [
+        "直接套用旧项目事实。",
+        "不区分评分点、章节草稿和待补资料。",
+        "把生成内容当成正式投标定稿。"
+      ],
+      "bannedClaims": [
+        "中标概率",
+        "评标结论",
+        "最终投标决策"
+      ]
+    },
+    {
+      "id": "scheme-writing-expert-comments",
+      "skillId": "aios-scheme-write",
+      "promptPath": "skills/aios-scheme-write/prompts/basic-prompt.md",
+      "scenario": "基于方案初稿、历史方案、工程概况和专家意见生成专项施工方案章节工作母版",
+      "sourceSignals": [
+        "source-shape:special construction scheme writing task with expert comments and historical scheme material",
+        "source-shape:scheme draft needs rewriting into a markdown working draft",
+        "boundary-shape:calculation correctness and approval outcomes require professional review"
+      ],
+      "advisoryComparison": [
+        "普通提示词容易把专家意见改成完整正文，并默认计算、附图和审批已经满足。",
+        "写作型基础提示词要求保留危险源、计算书缺口、专家意见回查和审核门禁。"
+      ],
+      "weakPrompt": "请根据方案初稿、专家意见和历史方案，帮我写一版专项施工方案。",
+      "inputSummary": [
+        "输入包含工程概况、专项施工方案初稿、历史方案素材和专家意见。",
+        "历史方案中的旧项目条件、设备参数和验收口径不能自动套入新项目。",
+        "缺少计算书、附图、危大工程审批或专业复核时，只能输出待补占位。"
+      ],
+      "sampleInput": [
+        "资料类型：专项施工方案章节生成任务。",
+        "工程概况摘要：本项目需要补充关键工序、危险源控制、检查验收、应急措施和交底要点。",
+        "专家意见摘要：需回查危险源识别、计算书目录、附图一致性、验收要求和应急措施。",
+        "历史方案片段：旧项目包含设备型号、荷载参数、平面布置和验收流程。",
+        "资料缺口：未提供本项目计算书全文、附图、设备厂家资料、专项审批记录和安全负责人确认意见。"
+      ],
+      "requiredPromptTerms": [
+        "Markdown",
+        "复用级别",
+        "方案章节结构",
+        "危险源和控制措施占位表",
+        "aios-construction-scheme"
+      ],
+      "expectedOutputShape": [
+        "先列资料来源和写作 brief。",
+        "把历史方案按可复用改写、仅参考、不可套用、需人工确认分级。",
+        "输出方案章节结构、Markdown 章节初稿、危险源占位表和待补资料。",
+        "不确认计算正确或审批通过，并交给审核 Skill 复核。"
+      ],
+      "expectedStrongSections": [
+        "资料来源清单",
+        "写作 brief",
+        "历史素材匹配表",
+        "方案章节结构",
+        "方案章节初稿 / 改写稿",
+        "危险源和控制措施占位表",
+        "审核门禁交接"
+      ],
+      "weakFailureModes": [
+        "直接套用旧项目参数和设备配置。",
+        "忽略专家意见、危险源和计算书缺口。",
+        "把方案初稿写成已经通过审批的定稿。"
+      ],
+      "bannedClaims": [
+        "方案合格",
+        "计算正确",
+        "专家论证通过",
+        "审批通过"
+      ]
+    }
+  ]
+}

package/prompts/evaluations/engineering-document-writing-scorecard.json

@@ -0,0 +1,108 @@
+{
+  "schema": 1,
+  "name": "engineering-document-writing-scorecard",
+  "version": "0.1",
+  "fixture": "prompts/evaluations/engineering-document-writing-fixtures.json",
+  "dataBoundary": "Structured comparison derived from de-identified engineering document writing fixtures. Do not include customer names, contacts, project names, exact amounts, or raw source documents.",
+  "minimumWeightedDelta": 1.2,
+  "criteria": [
+    {
+      "id": "source_traceability",
+      "weight": 20,
+      "description": "Keeps source list, provenance comments, and missing evidence visible instead of writing unsupported facts."
+    },
+    {
+      "id": "material_reuse",
+      "weight": 18,
+      "description": "Classifies historical material reuse and prevents old project facts, parameters, people, equipment, dates, or claims from being copied."
+    },
+    {
+      "id": "draft_actionability",
+      "weight": 17,
+      "description": "Produces a Markdown working draft and tables that engineering staff can continue editing."
+    },
+    {
+      "id": "boundary_safety",
+      "weight": 18,
+      "description": "Avoids tender decisions, approval outcomes, calculation correctness, compliance conclusions, and other over-claims."
+    },
+    {
+      "id": "review_gate",
+      "weight": 17,
+      "description": "Hands the generated draft to the correct review Skill with concrete issues to verify."
+    },
+    {
+      "id": "handoff_readability",
+      "weight": 10,
+      "description": "Keeps the output readable for project, tender, technical, safety, and reviewer roles."
+    }
+  ],
+  "cases": [
+    {
+      "caseId": "tender-writing-historical-material-reuse",
+      "winner": "basic",
+      "decisionBasis": "The basic writing prompt turns loose source material into a traceable Markdown workbench with scoring-point mapping, material reuse judgment, blocked old facts, and a tender review gate.",
+      "observedWeakFailures": [
+        "直接套用旧项目事实。",
+        "不区分评分点、章节草稿和待补资料。",
+        "把生成内容当成正式投标定稿。"
+      ],
+      "basicPromptGains": [
+        "Keeps source list, writing brief, and historical material matching table before drafting.",
+        "Requires reuse levels and removes old project facts.",
+        "Hands scoring point response and bid-risk review to aios-commercial-tender."
+      ],
+      "weakScores": {
+        "source_traceability": 2,
+        "material_reuse": 1,
+        "draft_actionability": 3,
+        "boundary_safety": 2,
+        "review_gate": 1,
+        "handoff_readability": 3
+      },
+      "basicScores": {
+        "source_traceability": 5,
+        "material_reuse": 5,
+        "draft_actionability": 4,
+        "boundary_safety": 5,
+        "review_gate": 5,
+        "handoff_readability": 4
+      }
+    },
+    {
+      "caseId": "scheme-writing-expert-comments",
+      "winner": "basic",
+      "decisionBasis": "The basic writing prompt keeps calculations, drawings, expert comments, hazard controls, and approvals as evidence-bound placeholders, then routes the draft through the scheme review gate.",
+      "observedWeakFailures": [
+        "直接套用旧项目参数和设备配置。",
+        "忽略专家意见、危险源和计算书缺口。",
+        "把方案初稿写成已经通过审批的定稿。"
+      ],
+      "basicPromptGains": [
+        "Separates reusable wording from project-specific equipment and calculation facts.",
+        "Forces hazard-control placeholders and missing calculation evidence into the draft.",
+        "Hands calculation gaps, expert comments, and safety review to aios-construction-scheme."
+      ],
+      "weakScores": {
+        "source_traceability": 2,
+        "material_reuse": 1,
+        "draft_actionability": 3,
+        "boundary_safety": 1,
+        "review_gate": 1,
+        "handoff_readability": 3
+      },
+      "basicScores": {
+        "source_traceability": 5,
+        "material_reuse": 5,
+        "draft_actionability": 4,
+        "boundary_safety": 5,
+        "review_gate": 5,
+        "handoff_readability": 4
+      }
+    }
+  ],
+  "overallDecision": {
+    "winner": "basic",
+    "notAClaim": "This scorecard validates deterministic fixture design and prompt guardrails; it does not prove real model performance or customer-ready document quality without skill-runtime output review."
+  }
+}

package/prompts/evaluations/engineering-knowledge-pack-scorecard.json

@@ -0,0 +1,57 @@
+{
+  "schema": 1,
+  "name": "engineering-knowledge-pack-scorecard",
+  "version": "0.1",
+  "dataBoundary": "Quality gate for synthetic or authorized AIOS Knowledge Packs. It evaluates source governance, traceability, runtime lookup behavior, and eval coverage; it does not certify real regulatory compliance.",
+  "criteria": [
+    {
+      "id": "source_authority",
+      "weight": 15,
+      "description": "Every source has authorization, version, hash or locator, review status, and clear boundary text."
+    },
+    {
+      "id": "version_applicability",
+      "weight": 15,
+      "description": "Standards and clauses preserve region, discipline, sourceVersion, effective status, and inapplicable version behavior."
+    },
+    {
+      "id": "clause_structure",
+      "weight": 15,
+      "description": "Clauses are decomposed into objects, conditions, requirement, exceptions, applicability, evidence, and review status."
+    },
+    {
+      "id": "graph_traceability",
+      "weight": 12,
+      "description": "Graph entities and relations retain source references and confidence, and never rely on unmarked model inference."
+    },
+    {
+      "id": "runtime_lookup",
+      "weight": 18,
+      "description": "Reference Runtime returns status, citations, applicability, sourceVersion, and notes through the knowledge.norm_lookup contract."
+    },
+    {
+      "id": "eval_gate",
+      "weight": 15,
+      "description": "Eval questions cover answerable, unanswerable, conflict, version mismatch, and need-context cases."
+    },
+    {
+      "id": "human_review_boundary",
+      "weight": 10,
+      "description": "The pack keeps draft/reviewed/blocked/published states explicit and does not replace formal expert review."
+    }
+  ],
+  "sampleCase": {
+    "packId": "scheme-review-reference",
+    "minimumPassingWeightedScore": 4.5,
+    "scores": {
+      "source_authority": 5,
+      "version_applicability": 5,
+      "clause_structure": 5,
+      "graph_traceability": 5,
+      "runtime_lookup": 5,
+      "eval_gate": 5,
+      "human_review_boundary": 5
+    },
+    "decisionBasis": "The sample pack is synthetic but complete: it declares source boundaries, compiles traceable clauses, exposes a local lookup runtime, and passes all deterministic eval cases."
+  }
+}

package/prompts/evaluations/skill-runtime/README.md

@@ -0,0 +1,15 @@
+# Skill Runtime Evidence
+
+本目录保存真实宿主触发 Skill 的证据归档。
+
+`skill-runtime` 只能来自宿主工具实际加载并触发 `$aios-*` Skill 后的原始输出，不接受把 `SKILL.md` 或 `basic-prompt.md` 当普通 prompt 粘贴运行的结果。
+
+每次归档至少包含：
+
+- 宿主名称、版本或可用性检查。
+- AIOS 版本或 commit。
+- 输入样板路径。
+- 触发短指令。
+- 是否能确认真实 Skill 触发。
+- 原始输出全文路径。
+- 若无法确认，写明 blocker，不得伪造 raw output。

package/prompts/evaluations/skill-runtime/raw/codex-scheme-writing-sample.md

@@ -0,0 +1,111 @@
+# Codex Raw Output: aios-scheme-write
+
+Host: Codex desktop current thread
+Captured at: 2026-06-19
+Skill loaded from: `C:\Users\white\.codex\skills\aios-scheme-write\SKILL.md`
+Triggered skill: `aios-scheme-write`
+Expected gate skill: `aios-construction-scheme`
+Sample workspace: `templates/document-writing-samples/scheme`
+Trigger prompt:
+
+```text
+$aios-scheme-write 请基于该工作台生成或改写 draft.md，并保留交给 aios-construction-scheme 的审核门禁。
+```
+
+## 1. 写作任务判断
+
+这是专项施工方案章节生成 / 改写任务，输入包含工程概况摘要、专家意见摘要、历史方案素材和素材复用索引，适用 `aios-scheme-write`。本输出只形成 Markdown 工作母版，不输出“方案合格 / 不合格 / 已通过审查”，也不替代总工、专家论证、危大工程审批、结构计算、监理或安全负责人。
+
+## 2. 资料来源清单
+
+| 序号 | 资料 | 版本 / 定位 | 当前用途 | 缺口 |
+|---:|---|---|---|---|
+| 1 | 工程概况摘要 | 示例版，方案概况 | 形成关键工序、危险源、验收和交底章节结构 | 缺少图纸编号 |
+| 2 | 专家意见摘要 | 示例版，修改意见 | 回查危险源、计算书、附图、验收和交底要点 | 缺少专家签认页 |
+| 3 | 历史方案素材 | 示例版，设备参数、验收流程 | 判断可复用章节表达 | 旧项目设备参数、图号和验收口径不得直接套用 |
+
+## 3. 写作 brief
+
+| 项目 | 内容 |
+|---|---|
+| 目标章节 | 关键工序、危险源控制、检查验收和交底要点 |
+| 输出格式 | Markdown 工作母版 |
+| 写作口径 | 只使用已提供资料，缺证据处保留 `待补` / `需核验` / `转人工复核` |
+| 禁止结论 | 不输出方案合格、计算正确、专家论证通过、审批通过或现场安全满足 |
+| 人工复核 | 技术负责人、总工、安全负责人、质量负责人、资料负责人 |
+
+## 4. 历史素材匹配表
+
+| 素材来源 | 原章节 / 原段落 | 本次可用位置 | 复用级别 | 需要删除的旧项目事实 | 需要补充的本项目证据 | 建议复核岗位 |
+|---|---|---|---|---|---|---|
+| 历史方案素材 | 关键工序描述 | 工艺流程 | 可复用改写 | 旧项目部位名称 | 本项目工序划分 | 技术负责人 |
+| 历史方案素材 | 设备参数 | 计算和设备能力 | 不可套用 | 设备型号、荷载参数 | 本项目设备资料和计算书 | 总工 |
+| 历史方案素材 | 验收流程 | 检查验收章节 | 仅参考 | 旧项目验收口径 | 本项目规范依据和验收要求 | 质量负责人 |
+
+## 5. 方案章节结构
+
+1. 工程概况与编制边界
+2. 施工准备与关键工序
+3. 危险源识别与控制措施
+4. 检查验收与资料留痕
+5. 应急措施
+6. 班组交底要点
+7. 待补资料和人工复核清单
+
+## 6. 施工方案 Markdown 初稿 / 改写稿
+
+### 工程概况与编制边界
+
+本方案章节基于已提供的工程概况摘要、专家意见摘要和历史方案素材进行工作母版整理。当前资料能够支持章节框架、回查清单和通用管理表达，但不能支持具体图纸编号、设备参数、荷载取值、计算结论或审批状态。
+
+图纸编号、计算书全文、设备厂家资料、专项审批记录和安全负责人确认意见均应作为 `待补` 项列入后续复核。
+
+### 关键工序
+
+关键工序章节建议围绕施工准备、工艺流程、过程控制、检查验收和应急处置展开。历史方案中的工艺流程表达可参考并改写，但必须删除旧项目部位名称、设备型号、参数和图号。
+
+建议正文写法：
+
+> 施工前应完成技术交底、作业条件确认、设备材料进场核查和安全防护准备。施工过程中按工序划分设置质量控制点、安全控制点和验收节点。具体工序划分、图纸编号、设备参数和检查频次待本项目资料补齐后确认。
+
+### 危险源控制
+
+危险源控制章节应把专家意见中的“危险源识别覆盖关键工序”拆成清单。每项危险源需绑定已有资料依据、控制措施、检查频次、责任岗位和待补资料；缺少本项目资料时不得写成已确认安全措施。
+
+## 7. 危险源和控制措施占位表
+
+| 序号 | 危险源 / 回查点 | 当前依据 | 控制措施草稿 | 待补资料 | 建议复核岗位 |
+|---:|---|---|---|---|---|
+| 1 | 关键工序危险源识别 | 专家意见摘要 | 按工序拆分危险源、控制措施和责任岗位 | 本项目工序划分 | 安全负责人 |
+| 2 | 设备能力和参数 | 历史方案素材提示有设备参数 | 不写具体参数，仅列复核项 | 设备厂家资料、计算书全文 | 总工 |
+| 3 | 检查验收 | 历史方案验收流程仅参考 | 建立检查验收清单和资料留痕 | 本项目规范依据、验收要求 | 质量负责人 |
+
+## 8. 待补资料和人工复核清单
+
+| 序号 | 待补内容 | 影响章节 | 建议复核岗位 |
+|---:|---|---|---|
+| 1 | 当前图纸编号 | 工程概况、附图引用 | 技术负责人 |
+| 2 | 计算书全文 | 设备能力和安全系数 | 总工 |
+| 3 | 设备厂家资料 | 机械设备参数 | 安全负责人 |
+| 4 | 专家签认页和专项审批记录 | 专家意见回查、审批状态说明 | 项目负责人 / 资料负责人 |
+
+## 9. 审核门禁交接
+
+请将本 `draft.md` 交给 `aios-construction-scheme` 复核，重点检查：
+
+- 危险源、控制措施、交底要点是否有方案原文或本项目条件依据。
+- 规范、地方规程、危大工程判断是否缺少工具、项目资料或专家证据。
+- 计算书、荷载、稳定、变形、设备能力和安全系数是否需要 `aios-structural` 或项目计算书复核。
+- 是否把历史方案中的旧项目地点、参数、工艺、设备、图号或工期写成未经证实的本次事实。
+- 是否存在方案合格、计算正确、专家论证通过、审批通过或现场安全满足等越权结论。
+
+## 10. 输出自检
+
+| 检查项 | 结果 |
+|---|---|
+| 是否包含资料来源清单 | 是 |
+| 是否保留素材复用级别 | 是 |
+| 是否保留 `待补` / `需核验` / `转人工复核` | 是 |
+| 是否避免输出合格、计算正确或审批通过结论 | 是 |
+| 是否明确 `aios-construction-scheme` 审核门禁 | 是 |
+| 是否把 AI 草稿标为正式定稿 | 否 |